Hive

Long-term memory and multi-agent orchestration for Claude Code.

Hive is a Claude Code plugin that gives Claude two superpowers it doesn't have by default:

1. Persistent memory across all projects and sessions — stored locally in SQLite, never sent anywhere.
2. Specialist agents (Planner, Coder, Debugger, Reviewer) running in parallel, coordinated by an orchestrator that loads relevant memories before dispatching work.

---

Why Hive?

Problem	Hive's solution
Claude forgets everything between sessions	SQLite FTS5 memory, survives restarts
One agent does everything mediocrely	Specialists: plan → code → debug → review
Memory systems need Chroma, LangChain, etc.	Zero dependencies: just sqlite3 (built into macOS)
Compression costs tokens over time	claude-haiku compresses old memories cheaply
Memory leaks across unrelated projects	Per-project scoping + global fallback

---

Install

git clone https://github.com/avinashcheby/Hive-Claude-Code-Skill
cd Hive-Claude-Code-Skill
bash install.sh

install.sh:

• Copies scripts to ~/.hive/scripts/
• Copies skills to ~/.claude/skills/hive*/ (auto-discovered by Claude Code)
• Initializes the SQLite database at ~/.hive/memory.db
• Cleans up any previous plugin-registry install

Then restart Claude Code. The /hive skills will be available automatically — no flags needed.

Requirements: macOS (sqlite3 built-in) or Linux with sqlite3 + FTS5. No Python, Node, or npm.

---

Commands

/hive <task>

The main command. Give it any task and Hive will:

1. Load relevant memories from past sessions
2. Route to appropriate specialist agents (in parallel)
3. Synthesize the results
4. Save learnings back to memory

/hive build a REST API for user authentication with JWT
/hive refactor the database layer to use connection pooling
/hive the login endpoint returns 500 when the email has a + sign

/hive-memory [search|delete|list]

Browse and manage your memory store.

/hive-memory search postgres
/hive-memory list error
/hive-memory delete 42

/hive-status

Show memory counts, recent sessions, DB size, and compression history.

/hive-compress

Compress old, low-importance memories using claude-haiku. Run this after large sessions to keep the DB lean.

---

How Memory Works

Memories are stored in ~/.hive/memory.db (SQLite with FTS5 full-text search).

Five memory types:

Type	Example
fact	"Project uses PostgreSQL 15 with pgvector"
decision	"Chose Redis over Memcached because TTL semantics are simpler"
pattern	"All API handlers return (Response, StatusCode), never panic"
error	"SQLite 'database is locked' when WAL mode is off"
preference	"Prefers table-driven tests, dislikes mocks"

Ranking formula:

score = (importance/10) × (1 / (1 + days_since_access)) × (1 + access_count)

High importance + recent + frequently used = top of recall results.

Compression: Memories with importance ≤ 3, never accessed, older than 30 days are summarized by claude-haiku into a single pattern entry. Tune with env vars:

export HIVE_COMPRESS_AGE=30               # days threshold
export HIVE_COMPRESS_MAX_IMPORTANCE=3     # max importance to compress
export HIVE_COMPRESS_BATCH=20             # memories per batch

---

How Agents Work

Hive has four specialist agents:

Agent	Role	When used
Planner	Decomposes task → steps, files, risks	Multi-step tasks, architecture
Coder	Implements code + tests	Any implementation work
Debugger	Root cause analysis + minimal fix	Error messages, test failures
Reviewer	Security, correctness, convention	Code touching APIs, auth, storage

The orchestrator (/hive) decides which agents to launch based on the task, launches them in parallel, then synthesizes the results. It never launches agents whose output it won't use.

---

Configuration

Env var	Default	Description
HIVE_DB	~/.hive/memory.db	Override DB location
HIVE_HOME	~/.hive	Override DB directory
HIVE_COMPRESS_AGE	30	Days before compression eligibility
HIVE_COMPRESS_MAX_IMPORTANCE	3	Max importance level to compress
HIVE_COMPRESS_BATCH	20	Memories per compression batch

---

Running Tests

bash tests/run_all.sh

Tests use isolated temp DBs and do not make network calls.

---

Contributing

See CLAUDE.md for contributor conventions.

Key rules:

• No Python/Node dependencies — bash + sqlite3 only
• Bash 3.2 compatible (macOS built-in)
• set -euo pipefail on every script
• Max 50 lines per function, 500 lines per file
• Every script in scripts/ gets a test in tests/

---

License

MIT